Alibaba's CPI-Bench Splits Image-Editing Evaluation Into General, Practical and Intelligent Subsets — and Claims the Closest Alignment to the Arena Image Edit Leaderboard
CPI-Bench (arXiv 2608.14546, submitted Aug 14; 12 upvotes on today's HuggingFace Daily Papers, 20 authors from Alibaba) restructures image-editing evaluation into three subsets: CPI-General-Bench for diverse tasks including multi-image editing, CPI-Practical-Bench for real user applications, and CPI-Intelligent-Bench for reasoning-based edits. Its central claim is correlational rather than architectural — that it 'achieves the highest alignment with the Arena Image Edit Leaderboard' among existing benchmarks, making it a cheap offline proxy for human preference. That is the useful property if you are picking an editing model without running a human eval, though a single self-reported alignment number from the benchmark's own authors warrants independent replication before you rank models by it.
↳ Follow the thread